Self-Driving Car Engineer Nanodegree¶
Deep Learning¶
Project: Build a Traffic Sign Recognition Classifier¶
In this notebook, a template is provided for you to implement your functionality in stages, which is required to successfully complete this project. If additional code is required that cannot be included in the notebook, be sure that the Python code is successfully imported and included in your submission if necessary.
Note: Once you have completed all of the code implementations, you need to finalize your work by exporting the iPython Notebook as an HTML document. Before exporting the notebook to html, all of the code cells need to have been run so that reviewers can see the final implementation and output. You can then export the notebook by using the menu above and navigating to \n", "File -> Download as -> HTML (.html). Include the finished document along with this notebook as your submission.
In addition to implementing code, there is a writeup to complete. The writeup should be completed in a separate file, which can be either a markdown file or a pdf document. There is a write up template that can be used to guide the writing process. Completing the code template and writeup template will cover all of the rubric points for this project.
The rubric contains "Stand Out Suggestions" for enhancing the project beyond the minimum requirements. The stand out suggestions are optional. If you decide to pursue the "stand out suggestions", you can include the code in this Ipython notebook and also discuss the results in the writeup file.
Note: Code and Markdown cells can be executed using the Shift + Enter keyboard shortcut. In addition, Markdown cells can be edited by typically double-clicking the cell to enter edit mode.
Step 0: Load The Data¶
## Load pickled data
import pickle
training_file = 'data/train.p'
validation_file= 'data/valid.p'
testing_file = 'data/test.p'
with open(training_file, mode='rb') as f:
train = pickle.load(f)
with open(validation_file, mode='rb') as f:
valid = pickle.load(f)
with open(testing_file, mode='rb') as f:
test = pickle.load(f)
X_train, y_train = train['features'], train['labels']
X_valid, y_valid = valid['features'], valid['labels']
X_test, y_test = test['features'], test['labels']
Step 1: Dataset Summary & Exploration¶
The pickled data is a dictionary with 4 key/value pairs:
'features'is a 4D array containing raw pixel data of the traffic sign images, (num examples, width, height, channels).'labels'is a 1D array containing the label/class id of the traffic sign. The filesignnames.csvcontains id -> name mappings for each id.'sizes'is a list containing tuples, (width, height) representing the original width and height the image.'coords'is a list containing tuples, (x1, y1, x2, y2) representing coordinates of a bounding box around the sign in the image. THESE COORDINATES ASSUME THE ORIGINAL IMAGE. THE PICKLED DATA CONTAINS RESIZED VERSIONS (32 by 32) OF THESE IMAGES
Complete the basic data summary below. Use python, numpy and/or pandas methods to calculate the data summary rather than hard coding the results. For example, the pandas shape method might be useful for calculating some of the summary results.
Provide a Basic Summary of the Data Set Using Python, Numpy and/or Pandas¶
### Use python, pandas or numpy methods rather than hard coding the results
import numpy as np
# TODO: Number of training examples
n_train = len(X_train)
# TODO: Number of validation examples
n_valid = len(X_valid)
# TODO: Number of testing examples.
n_test = len(X_test)
# TODO: What's the shape of a traffic sign image?
image_shape = X_train.shape[1:]
# TODO: How many unique classes/labels there are in the dataset.
labels_index = np.unique(y_train)
n_labels = len(labels_index)
print("Number of training examples =", n_train)
print("Number of validation examples =", n_valid)
print("Number of testing examples =", n_test)
print("Image data shape =", image_shape)
print("Number of labels/classes =", n_labels)
Include an exploratory visualization of the dataset¶
Visualize the German Traffic Signs Dataset using the pickled file(s). This is open ended, suggestions include: plotting traffic sign images, plotting the count of each sign, etc.
The Matplotlib examples and gallery pages are a great resource for doing visualizations in Python.
NOTE: It's recommended you start with something simple first. If you wish to do more, come back to it after you've completed the rest of the sections. It can be interesting to look at the distribution of classes in the training, validation and test set. Is the distribution the same? Are there more examples of some classes than others?
## Show random image and label index
import random
import matplotlib.pyplot as plt
# Show visualizations in the notebook
%matplotlib inline
index = random.randint(0, len(X_train))
image = X_train[index].squeeze()
plt.figure(figsize=(1,1))
plt.imshow(image, cmap="gray")
print('label index: ', y_train[index])
Complete set of images and labels¶
## Create label dictionary {label:signname}
import csv
labels_dict = None
with open('signnames.csv', mode='r') as infile:
reader = csv.reader(infile)
next(reader, None)
labels_dict = {int(rows[0]):rows[1] for rows in reader}
print(labels_dict)
Original Images (before preprocessing)¶
## Display sample of ORIGINAL training images (BEFORE preprocessing)
fig = plt.figure(figsize=(22,35), tight_layout={'h_pad':5})
for i in range(n_labels):
image_key = np.where(y_train==i)
img = X_train[image_key[0][0]]
ax = fig.add_subplot(int(n_labels/4)+1,4,i+1)
ax.imshow(img, interpolation='none')
plt.title("[%02d] %s" % (i, labels_dict[y_train[image_key[0][0]]]))
plt.show()
## Display label distribution in a histogram
import pandas as pd
y_train_df = pd.DataFrame()
y_train_df['label'] = y_train
ax = y_train_df['label'].value_counts().plot(kind='barh', figsize = (10,10), title='Number of Samples per Class')
ax.set_yticklabels(list(map(lambda x: labels_dict[x], y_train_df['label'].value_counts().index.tolist())))
for i, v in enumerate(y_train_df['label'].value_counts()):
ax.text(v + 10, i - 0.25, str(v), color='black')
## Display labels in a table
import pandas as pd
labels_i = [[i, sum(y_train == i)] for i in range(len(np.unique(y_train)))]
data_pd = pd.read_csv('signnames.csv')
labels_pd = data_pd.set_index('ClassId')
labels_pd['Count'] = pd.Series(np.asarray(labels_i).T[1], index=np.asarray(labels_i).T[0])
labels_pd
Step 2: Design and Test a Model Architecture¶
Design and implement a deep learning model that learns to recognize traffic signs. Train and test your model on the German Traffic Sign Dataset.
The LeNet-5 implementation shown in the classroom at the end of the CNN lesson is a solid starting point. You'll have to change the number of classes and possibly the preprocessing, but aside from that it's plug and play!
With the LeNet-5 solution from the lecture, you should expect a validation set accuracy of about 0.89. To meet specifications, the validation set accuracy will need to be at least 0.93. It is possible to get an even higher accuracy, but 0.93 is the minimum for a successful project submission.
There are various aspects to consider when thinking about this problem:
- Neural network architecture (is the network over or underfitting?)
- Play around preprocessing techniques (normalization, rgb to grayscale, etc)
- Number of examples per label (some have more than others).
- Generate fake data.
Here is an example of a published baseline model on this problem. It's not required to be familiar with the approach used in the paper but, it's good practice to try to read papers like these.
Pre-process the Data Set¶
Minimally, the image data should be normalized so that the data has mean zero and equal variance. For image data, (pixel - 128)/ 128 is a quick way to approximately normalize the data and can be used in this project.
Other pre-processing steps are optional. You can try different techniques to see if it improves performance.
Use the code cell (or multiple code cells, if necessary) to implement the first step of your project.
My approach:
- normalization will be handled within the network
- data augmentation will include grayscaling and affine transformations
Normalization¶
# Function that applies normalization and local contrast enhancement
from tqdm import tqdm
from skimage import color, data, exposure
def normalize(image_data):
'''Contrast Limited Adaptive Histogram Equalization (CLAHE). In addition to regular normalization,
this function provides local contrast enhancement -- i.e., details of the image can be
enhanced even in regions that are darker or lighter than most of the image.
http://scikit-image.org/docs/dev/api/skimage.exposure.html#skimage.exposure.equalize_adapthist
'''
norm = np.array([exposure.equalize_adapthist(image, clip_limit=0.1) for image in tqdm(image_data)])
return norm
# Run the normalization process
print('Normalizing Training Images...')
X_train_norm = normalize(X_train)
print('Normalizing Validation Images...')
X_valid_norm = normalize(X_valid)
print('\nNormalizing Test Images...')
X_test_norm = normalize(X_test)
print('\nNormaliztion Complete.')
## Display sample of NORMALIZED training images
fig = plt.figure(figsize=(22,35), tight_layout={'h_pad':5})
for i in range(n_labels):
image_key = np.where(y_train==i)
img = X_train_norm[image_key[0][0]]
ax = fig.add_subplot(int(n_labels/4)+1,4,i+1)
ax.imshow(img, interpolation='none')
plt.title("[%02d] %s" % (i, labels_dict[y_train[image_key[0][0]]]))
plt.show()
Augmentation¶
Generates batches of tensor image data that is augmented based on a chosen set of tranformation parameters (e.g. rotation, shift, shear, zoom).
## Apply affine and color transformations
import keras
from keras.preprocessing.image import ImageDataGenerator
# Transformations applied to RBG training images
datagen = keras.preprocessing.image.ImageDataGenerator(
zca_whitening=True,
rotation_range=10,
width_shift_range=0.1,
height_shift_range=0.1,
shear_range=0.15,
zoom_range=0.15,
channel_shift_range=0.1,
fill_mode='nearest',
horizontal_flip=False,
vertical_flip=False
)
## Create AUGMENTED training sets
from sklearn.utils import shuffle
max_images_per_class = 6000
X_train_aug = np.empty((0, 32, 32, 3))
y_train_aug = np.empty(0, dtype='uint8')
print('Augmenting Image Data...')
for i in tqdm(range(n_labels)):
index = [y_train==i]
i_class_images = X_train_norm[y_train==i]
i_class_y = y_train[y_train==i]
i_X_train_aug = np.copy(i_class_images)
i_y_train_aug = np.copy(i_class_y)
for X,y in datagen.flow(i_class_images, i_class_y, batch_size=len(i_class_y), seed=113+i*17):
i_X_train_aug = np.append(i_X_train_aug, X, axis=0)
i_y_train_aug = np.append(i_y_train_aug, y, axis=0)
if len(i_X_train_aug) >= max_images_per_class:
break
X_train_aug = np.append(X_train_aug, i_X_train_aug[:max_images_per_class], axis=0)
y_train_aug = np.append(y_train_aug, i_y_train_aug[:max_images_per_class], axis=0)
X_train_aug, y_train_aug = shuffle(X_train_aug, y_train_aug, random_state=113)
print('Augmentation Complete.')
print('new number of training images: ', X_train_aug.shape[0])
print('verify image shape: ', X_train_aug.shape[1:])
## Display sample of the AUGMENTED training images, 1 image for each class
fig = plt.figure(figsize=(22,35), tight_layout={'h_pad':5})
for i in range(n_labels):
image_key = np.where(y_train_aug==i)
img = X_train_aug[image_key[0][0]]
ax = fig.add_subplot(int(n_labels/4)+1,4,i+1)
ax.imshow(img, interpolation='none')
plt.title("[%02d] %s" % (i, labels_dict[y_train_aug[image_key[0][0]]]))
plt.axis('off')
plt.show()
# Create image dictionary {label:images}
train_images_dict = dict()
for i, (image, label) in enumerate(zip(X_train_aug, y_train_aug)):
if label not in train_images_dict:
train_images_dict[label] = []
train_images_dict[label].append(image)
# Verify that number of images in dict is correct
sum(len(v) for v in train_images_dict.values())
## Display a broader sample of AUGMENTED images, 12 from each class
import matplotlib
import matplotlib.pyplot as plt
%matplotlib inline
matplotlib.rcParams['figure.figsize'] = (20.0, 10.0)
images_labels = list(train_images_dict.keys())
for image_key in images_labels:
subplots = 12
fig, axes = plt.subplots(1,subplots)
images_total = len(train_images_dict[image_key])
print('[{}] {}'.format(image_key, labels_dict[image_key]))
offset = images_total // subplots
images = [train_images_dict[image_key][j*offset] for j in range(subplots)]
for i, (image, ax) in enumerate(zip(images, axes)):
ax.imshow(image.squeeze())
plt.tight_layout()
plt.subplots_adjust(top=0.85)
plt.show()
train_images_dict.clear()
Grayscaling¶
## Create grayscale training sets **3D**
from skimage import exposure, color, data
from sklearn.utils import shuffle
max_images_per_class = 3000
X_train_gray = np.empty((0, 32, 32, 1))
y_train_gray = np.empty(0, dtype='uint8')
print('Grayscaling Image Data...')
for i in tqdm(range(n_labels)):
index = [y_train_aug==i]
i_class_images = X_train_aug[y_train_aug==i]
i_class_y = y_train_aug[y_train_aug==i]
i_X_train_gray = []
i_y_train_gray = []
for ii in range(0, max_images_per_class):
image_gray = np.dot(i_class_images[ii], [[0.2125], [0.7154], [0.0721]])
i_X_train_gray.append(image_gray)
i_y_train_gray.append(i_class_y[ii])
X_train_gray = np.append(X_train_gray, np.asarray(i_X_train_gray), axis=0)
y_train_gray = np.append(y_train_gray, np.asarray(i_y_train_gray), axis=0)
X_train_gray, y_train_gray = shuffle(X_train_gray, y_train_gray)
print('\nGrayscaling Complete.')
print('number of grayscale images: ', X_train_gray.shape[0])
print('verify image shape: ', X_train_gray.shape[1:])
## Display original vs. normalized vs. grayscale versions
index = random.randint(0, len(X_train_norm))
image_orig = X_train[index].squeeze()
image_norm = X_train_norm[index].squeeze()
image_gray = np.dot(X_train_norm[index], [[0.2125], [0.7154], [0.0721]]).squeeze()
plt.figure(figsize=(5,5))
plt.subplot2grid((1, 3), (0, 0));
plt.axis('off')
plt.title('original')
plt.imshow(image_orig, cmap="gray")
plt.subplot2grid((1, 3), (0, 1));
plt.axis('off')
plt.title('normalized')
plt.imshow(image_norm, cmap="gray")
plt.subplot2grid((1, 3), (0, 2));
plt.axis('off')
plt.title('grayscale')
plt.imshow(image_gray, cmap="gray")
print('label index: ', y_train[index])
## Display sample of the GRAYSCALE training images, 1 image per class
fig = plt.figure(figsize=(22,35), tight_layout={'h_pad':5})
for i in range(n_labels):
image_key = np.where(y_train_gray==i)
img = X_train_gray[image_key[0][0]].squeeze()
ax = fig.add_subplot(int(n_labels/4)+1,4,i+1)
ax.imshow(img, cmap="gray")
plt.title("[%02d] %s" % (i, labels_dict[y_train_gray[image_key[0][0]]]))
plt.axis('off')
plt.show()
## Create training, validation, and testing sets for GRAYSCALE images
from sklearn.model_selection import train_test_split
X_train_gray, X_test_gray, y_train_gray, y_test_gray = \
train_test_split(X_train_gray, y_train_gray, test_size=0.2, random_state=12)
X_train_gray, X_valid_gray, y_train_gray, y_valid_gray = \
train_test_split(X_train_gray, y_train_gray, test_size=0.1, random_state=55)
## Verify GRAYSCALE data splits
gray_image_shape = X_train_gray.shape[1:]
gray_labels_index = np.unique(y_train_gray)
n_gray_labels = len(gray_labels_index)
print("Number of GRAYSCALE training examples =", len(X_train_gray))
print("Number of GRAYSCALE validation examples =", len(X_valid_gray))
print("Number of GRAYSCALE test examples =", len(X_test_gray))
print("Image data shape =", gray_image_shape)
print("Number of labels/classes =", n_gray_labels)
## Global variables and parameters
LOG_DIR = ".logs"
MODEL_DIR = ".models"
EPOCHS = 41
BATCH_SIZE = 128
lr = 0.001 # learning rate
decay= 0.0003 # for L2 reg
mu = 0 # normalized mean
sigma = 0.1 # normalized stdev
Operations¶
The series of functions below are designed to make the model more modular. This reduces the amount of hard-coding and makes it much easier to experiment with different model architectures.
# Creates convolutional layer
def conv_layer(input, n_kernels=1, kernel_size=(1, 1), strides=[1, 1, 1, 1], name="conv"):
H, W = kernel_size # filter height, width
C = input.get_shape().as_list()[3] # input depth
K = n_kernels # output depth
in_shape = [H, W, C, K]
print('{} filter shape: {}'.format(name, in_shape))
with tf.name_scope(name):
w = tf.Variable(tf.truncated_normal(in_shape, mean=mu, stddev=sigma), name="weights")
b = tf.Variable(tf.zeros(K), name="biases")
conv = tf.nn.conv2d(input, w, strides=strides, padding='SAME')
conv = tf.nn.bias_add(conv, b)
out_shape = conv.get_shape().as_list()
print('{} output shape: {}'.format(name, out_shape))
return conv
# Creates convolutional layer (for GRAYSCALE model only)
def conv_layer_gray(input, n_kernels=1, kernel_size=(1, 1), strides=[1, 1, 1, 1], name="conv"):
H, W = kernel_size # filter height, width
C = input.get_shape().as_list()[3] # input depth
K = n_kernels # output depth
in_shape = [H, W, C, K]
print('{} filter shape: {}'.format(name, in_shape))
with tf.name_scope(name):
w = tf.Variable(tf.truncated_normal(in_shape, mean=mu, stddev=sigma), name="weights")
b = tf.Variable(tf.zeros(K), name="biases")
conv = tf.nn.conv2d(input, w, strides=strides, padding='SAME')
conv = tf.nn.bias_add(conv, b)
out_shape = conv.get_shape().as_list()
print('{} output shape: {}'.format(name, out_shape))
return conv
# Creates fully connected layer
def fc_layer(input, n_inputs, n_outputs, name="fc"):
shape = input.get_shape().as_list()
print('{} shape: inputs={} outputs={}'.format(name, shape[1], n_outputs))
with tf.name_scope(name):
w = tf.Variable(tf.truncated_normal([n_inputs, n_outputs], mean=mu, stddev=sigma), name="weights")
b = tf.Variable(tf.zeros(n_outputs), name="biases")
act = tf.matmul(input, w) + b
return act
# Applies max pooling
def max_pool(input, kernel_size=(1,1), strides=[1,1,1,1], name="maxpool"):
H = kernel_size[0]
W = kernel_size[1]
print('{}: {}x{} / {}'.format(name, H, W, strides[1]))
with tf.name_scope(name):
pool = tf.nn.max_pool(input, ksize=[1, H, W, 1], strides=strides, padding='VALID', name=name)
return pool
Modified LeNet Models¶
# Model for AUGMENTED images
def LeNet_5(x, keep_prob):
with tf.name_scope("conv_layers"):
conv = tf.nn.relu(conv_layer(x, n_kernels=3, kernel_size=(1, 1), strides=[1, 1, 1, 1], name="conv_0"))
conv = tf.nn.relu(conv_layer(conv, n_kernels=8, kernel_size=(5, 5), strides=[1, 1, 1, 1], name="conv_1"))
conv = tf.nn.relu(conv_layer(conv, n_kernels=16, kernel_size=(5, 5), strides=[1, 1, 1, 1], name="conv_2"))
conv = max_pool(conv, kernel_size=(2,2), strides=[1,2,2,1])
conv = tf.nn.relu(conv_layer(conv, n_kernels=16, kernel_size=(5, 5), strides=[1, 1, 1, 1], name="conv_3"))
conv = tf.nn.relu(conv_layer(conv, n_kernels=32, kernel_size=(5, 5), strides=[1, 1, 1, 1], name="conv_4"))
conv = max_pool(conv, kernel_size=(2,2), strides=[1,2,2,1])
conv = tf.nn.relu(conv_layer(conv, n_kernels=32, kernel_size=(5, 5), strides=[1, 1, 1, 1], name="conv_5"))
conv = tf.nn.relu(conv_layer(conv, n_kernels=64, kernel_size=(5, 5), strides=[1, 1, 1, 1], name="conv_6"))
conv = max_pool(conv, kernel_size=(2,2), strides=[1,2,2,1])
flat = tf.contrib.layers.flatten(conv) # =[batch_size, k_features]
flat_shape = flat.get_shape().as_list()
k_features = flat_shape[1]
print('flat layer dim: ', flat_shape)
with tf.name_scope("fc_layers"):
fc = fc_layer(flat, n_inputs=k_features, n_outputs=1024, name="fc_1")
fc = tf.nn.relu(fc)
fc = tf.nn.dropout(fc, keep_prob)
fc = fc_layer(fc, n_inputs=1024, n_outputs=512, name="fc_2")
fc = tf.nn.relu(fc)
fc = tf.nn.dropout(fc, keep_prob)
logits = fc_layer(fc, n_inputs=512, n_outputs=43, name="output")
return logits
# Model for GRAYSCALE images
def LeNet_5_gray(x):
with tf.name_scope("conv_layers_gray"):
conv = tf.nn.relu(conv_layer_gray(x, n_kernels=1, kernel_size=(1, 1), strides=[1, 1, 1, 1], name="conv_0"))
conv = tf.nn.relu(conv_layer_gray(conv, n_kernels=6, kernel_size=(5, 5), strides=[1, 1, 1, 1], name="conv_1"))
conv = tf.nn.relu(conv_layer_gray(conv, n_kernels=16, kernel_size=(5, 5), strides=[1, 1, 1, 1], name="conv_2"))
conv = max_pool(conv, kernel_size=(2,2), strides=[1,2,2,1])
conv = tf.nn.relu(conv_layer_gray(conv, n_kernels=6, kernel_size=(5, 5), strides=[1, 1, 1, 1], name="conv_3"))
conv = tf.nn.relu(conv_layer_gray(conv, n_kernels=16, kernel_size=(5, 5), strides=[1, 1, 1, 1], name="conv_4"))
conv = max_pool(conv, kernel_size=(2,2), strides=[1,2,2,1])
flat = tf.contrib.layers.flatten(conv) # =[batch_size, k_features]
flat_shape = flat.get_shape().as_list()
k_features = flat_shape[1]
print('flat layer dim: ', flat_shape)
with tf.name_scope("fc_layers_gray"):
fc = fc_layer(flat, n_inputs=k_features, n_outputs=1024, name="fc_1")
fc = tf.nn.relu(fc)
fc = tf.nn.dropout(fc, keep_prob)
fc = fc_layer(fc, n_inputs=1024, n_outputs=512, name="fc_2")
fc = tf.nn.relu(fc)
fc = tf.nn.dropout(fc, keep_prob)
logits = fc_layer(fc, n_inputs=512, n_outputs=43, name="output")
return logits
Train, Validate and Test each Model¶
A validation set can be used to assess how well the model is performing. A low accuracy on the training and validation sets imply underfitting. A high accuracy on the training set but low accuracy on the validation set implies overfitting.
Training & Loss Functions¶
# cross entropy
def cross_entropy(logits, one_hot_y):
with tf.name_scope("xent"):
xent = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=logits, labels=one_hot_y))
tf.summary.scalar("xent", xent)
return xent
# training
def train_step(lr, xent):
with tf.name_scope("train"):
train_step = tf.train.AdamOptimizer(learning_rate=lr).minimize(xent + l2)
return train_step
# accuracy
def accuracy(logits, one_hot_y):
with tf.name_scope("accuracy"):
pred = tf.equal(tf.argmax(logits, 1), tf.argmax(one_hot_y, 1))
acc = tf.reduce_mean(tf.cast(pred, tf.float32))
tf.summary.scalar("accuracy", acc)
return acc
# L2 regularization
def l2_reg(decay):
train_vars = tf.trainable_variables()
with tf.name_scope("L2"):
l2 = tf.add_n([tf.nn.l2_loss(v) for v in train_vars if 'weight' in v.name]) * decay
return l2
# Calculates accuracy over entire input data set
def evaluate(x_data, y_data):
num_examples = len(x_data)
total_acc = 0
sess = tf.get_default_session()
for offset in range(0, num_examples, BATCH_SIZE):
batch_x, batch_y = x_data[offset:offset+BATCH_SIZE], y_data[offset:offset+BATCH_SIZE]
batch_acc = sess.run(acc, feed_dict={x: batch_x, y: batch_y, keep_prob: 1.0})
total_acc += (batch_acc * len(batch_x))
final_acc = total_acc / num_examples
return final_acc
Construct the Graph¶
## The graph for training on AUGMENTED images
import tensorflow as tf
tf.reset_default_graph()
x = tf.placeholder(tf.float32, (None, 32, 32, 3))
y = tf.placeholder(tf.int32, (None))
one_hot_y = tf.one_hot(y, 43)
keep_prob = tf.placeholder(tf.float32, (None))
print('Verify Network Structure:\n--------------')
logits = LeNet_5(x, keep_prob)
l2 = l2_reg(decay)
xent = cross_entropy(logits, one_hot_y)
train = train_step(lr, xent)
acc = accuracy(logits, one_hot_y)
## The graph for training GRAYSCALE images
import tensorflow as tf
tf.reset_default_graph()
x = tf.placeholder(tf.float32, (None, 32, 32, 1))
y = tf.placeholder(tf.int32, (None))
one_hot_y = tf.one_hot(y, 43)
keep_prob = tf.placeholder(tf.float32)
print('Verify Network Structure:\n--------------')
logits = LeNet_5_gray(x, keep_prob)
l2 = l2_reg(decay)
xent = cross_entropy(logits, one_hot_y)
train = train_step(lr, xent)
acc = accuracy(logits, one_hot_y)
Training¶
##### NEW GOLD MASTER #####
## Training and accuracy on AUGMENTED data set
'''This model produces the best results:
Validation Accuracy = 99.4%
Test Accuracy = 98.2%
'''
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
writer = tf.summary.FileWriter(LOG_DIR, graph=tf.get_default_graph())
saver = tf.train.Saver()
num_examples = len(X_train_aug)
print("Training...\n--------")
for i in range(EPOCHS):
step = 0
for offset in range(0, num_examples, BATCH_SIZE):
step += 1
start = offset
end = offset + BATCH_SIZE
batch_x, batch_y = X_train_aug[start:end], y_train_aug[start:end]
sess.run(train, feed_dict={x: batch_x, y: batch_y, keep_prob: 0.5})
if step % 400 == 0:
batch_acc = sess.run(acc, feed_dict={x: batch_x, y: batch_y, keep_prob: 1.0})
print("Epoch {}, Step {}".format(i+1, step))
print("Training Accuracy = {:.3f}\n".format(batch_acc))
# Calculate validation accuracy
valid_acc = evaluate(X_valid_norm, y_valid)
print("Validation Accuracy = {:.3f}\n".format(valid_acc))
print("--------")
if valid_acc > 0.9925:
break
# Calculate test accuracy (run only on final model)
test_acc = evaluate(X_test_norm, y_test)
print("\nTest Accuracy = {:.3f}\n".format(test_acc))
print("--------")
saver.save(sess, MODEL_DIR)
print("Model saved")
# Training and accuracy on GRAYSCALE data set
'''This model also produces good results:
Validation Accuracy = 95.8%
Test Accuracy = 94.5%
'''
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
writer = tf.summary.FileWriter(LOG_DIR, graph=tf.get_default_graph())
saver = tf.train.Saver()
num_examples = len(X_train_gray)
print("Training...\n--------")
for i in range(EPOCHS):
step = 0
for offset in range(0, num_examples, BATCH_SIZE):
step += 1
start = offset
end = offset + BATCH_SIZE
batch_x, batch_y = X_train_gray[start:end], y_train_gray[start:end]
sess.run(train, feed_dict={x: batch_x, y: batch_y})
if step % 400 == 0:
batch_acc = sess.run(acc, feed_dict={x: batch_x, y: batch_y})
print("Epoch {}, Step {}".format(i+1, step))
print("Training Accuracy = {:.3f}\n".format(batch_acc))
# Calculate validation accuracy
valid_accuracy = evaluate(X_valid_gray, y_valid_gray)
print("Validation Accuracy = {:.3f}\n".format(valid_accuracy))
print("--------")
# Calculate test accuracy (run only on final model)
keep_prob = 1.0
test_acc = evaluate(X_test_gray, y_test_gray)
print("\nTest Accuracy = {:.3f}\n".format(test_acc))
print("--------")
saver.save(sess, MODEL_DIR)
print("Model saved")
Step 3: Test a Model on New Images¶
To give yourself more insight into how your model is working, download at least five pictures of German traffic signs from the web and use your model to predict the traffic sign type.
You may find signnames.csv useful as it contains mappings from the class id (integer) to the actual sign name.
Load and Output the Images¶
# Import and display the set of NEW images
import os
import matplotlib.pyplot as plt
import matplotlib.image as mpimg
import numpy as np
%matplotlib inline
# show cropped images
fig = plt.figure(figsize=(20, 80))#, tight_layout={'h_pad':2})
i = 0
image_dir = 'images/new-signs'
for file in os.listdir(image_dir):
if '.jpg' in file:
ax = fig.add_subplot(15,4,i+1)
img_crop = mpimg.imread(image_dir + '/' + file)
plt.title('TEST IMAGE ------>', {'verticalalignment': 'baseline', 'horizontalalignment': 'center'})
plt.axis('off')
ax.set_xlim([0,32])
ax.set_ylim([0,90])
# ax.set_autoscaley_on(False)
plt.gca().invert_yaxis()
ax.imshow(img_crop, origin='upper')
i+=1
ax = fig.add_subplot(15,4,i+1)
img_orig = mpimg.imread(image_dir + '/originals/' + file)
plt.title('------> ORIGINAL', {'verticalalignment': 'baseline', 'horizontalalignment': 'right'})
plt.axis('off')
# ax.set_xlim([0,300])
# ax.set_ylim([0,200])
# ax.set_autoscaley_on(False)
# plt.gca().invert_yaxis()
ax.imshow(img_orig, origin='upper')
i+=1
plt.show()
Predict the Sign Type for Each Image¶
### Run the predictions here and use the model to output the prediction for each image.
### Make sure to pre-process the images with the same pre-processing pipeline used earlier.
### Feel free to use as many code cells as needed.
Analyze Performance¶
### Calculate the accuracy for these 5 new images.
### For example, if the model predicted 1 out of 5 signs correctly, it's 20% accurate on these new images.
## Preprocess the NEW images
import os
import numpy as np
import matplotlib.image as mpimg
# Create labels
dir = 'images/new-signs'
new_filenames = os.listdir(dir)
y_new = [int(n.partition("-")[0]) for n in new_filenames if '.jpg' in n]
print('{} new labels created:\n{}'.format(len(y_new), y_new))
# Convert images to np arrays
new_files = [dir + '/' + n for n in new_filenames if '.jpg' in n]
X_new = [mpimg.imread(f) for f in new_files]
print('\n{} new images with shape: {}'.format(len(X_new), X_new[0].shape))
# Normalize the images
print('\nNormalizing new images...')
X_new_norm = normalize(X_new)
print('\nNormalization complete.')
# Calculate predictions and accuracy for new images
import tensorflow as tf
with tf.Session() as sess:
saver = tf.train.Saver()
saver.restore(sess, MODEL_DIR)
sess = tf.get_default_session()
preds = sess.run(tf.nn.softmax(logits), feed_dict={x: X_new_norm, keep_prob: 1.0})
print("\nPredicted Classes:")
preds_list = list(np.argmax(preds,axis=1))
print(preds_list)
print("vs. Actual Classes:")
y_new_ints = [int(i) for i in y_new]
print(y_new_ints)
new_acc = evaluate(X_new_norm, y_new)
print("\nNew Image Accuracy = {:.3f}\n".format(new_acc))
Output Top 5 Softmax Probabilities For Each Image Found on the Web¶
For each of the new images, print out the model's softmax probabilities to show the certainty of the model's predictions (limit the output to the top 5 probabilities for each image). tf.nn.top_k could prove helpful here.
The example below demonstrates how tf.nn.top_k can be used to find the top k predictions for each image.
tf.nn.top_k will return the values and indices (class ids) of the top k predictions. So if k=3, for each sign, it'll return the 3 largest probabilities (out of a possible 43) and the correspoding class ids.
Take this numpy array as an example. The values in the array represent predictions. The array contains softmax probabilities for five candidate images with six possible classes. tk.nn.top_k is used to choose the three classes with the highest probability:
# (5, 6) array
a = np.array([[ 0.24879643, 0.07032244, 0.12641572, 0.34763842, 0.07893497,
0.12789202],
[ 0.28086119, 0.27569815, 0.08594638, 0.0178669 , 0.18063401,
0.15899337],
[ 0.26076848, 0.23664738, 0.08020603, 0.07001922, 0.1134371 ,
0.23892179],
[ 0.11943333, 0.29198961, 0.02605103, 0.26234032, 0.1351348 ,
0.16505091],
[ 0.09561176, 0.34396535, 0.0643941 , 0.16240774, 0.24206137,
0.09155967]])Running it through sess.run(tf.nn.top_k(tf.constant(a), k=3)) produces:
TopKV2(values=array([[ 0.34763842, 0.24879643, 0.12789202],
[ 0.28086119, 0.27569815, 0.18063401],
[ 0.26076848, 0.23892179, 0.23664738],
[ 0.29198961, 0.26234032, 0.16505091],
[ 0.34396535, 0.24206137, 0.16240774]]), indices=array([[3, 0, 5],
[0, 1, 4],
[0, 5, 1],
[1, 3, 5],
[1, 4, 3]], dtype=int32))Looking just at the first row we get [ 0.34763842, 0.24879643, 0.12789202], you can confirm these are the 3 largest probabilities in a. You'll also notice [3, 0, 5] are the corresponding indices.
### Print out the top five softmax probabilities for the predictions on the German traffic sign images found on the web.
### Feel free to use as many code cells as needed.
with tf.Session() as sess:
saver.restore(sess, MODEL_DIR)
sess = tf.get_default_session()
pred = tf.nn.softmax(logits)
k = 5
topK = tf.nn.top_k(pred, k)
topK_pred = sess.run([pred, topK], feed_dict={x: X_new_norm, keep_prob: 1.0})
correct_count = 0
print('\n\n New Input Image: Top {} Predictions:'.format(k))
print(" ___________________________ _______________________________")
for i in range(0, len(X_new)):
plt.figure(figsize = (10,1.75))
plt.subplot2grid((1, 3), (0, 0));
plt.title('[{}] {}'.format(y_new[i], labels_pd['SignName'][int(y_new[i])]), \
{'verticalalignment': 'baseline', 'horizontalalignment': 'left'})
plt.axis('off');
plt.subplot2grid((1, 3), (0, 1));
plt.imshow(X_new[i], interpolation='none');
plt.axis('off');
plt.subplot2grid((1, 3), (0, 2));
plt.barh(6-np.arange(k),topK_pred[1][0][i], align='center');
for i_label in range(k):
class_id = topK_pred[1][1][i][i_label]
sign_name = labels_pd['SignName'][topK_pred[1][1][i][i_label]]
pred_i_label = topK_pred[1][0][i][i_label]
plt.text(pred_i_label+.02,6-i_label-.25, "[{}] {} {:.0%}".format(class_id, sign_name, pred_i_label));
pred_0 = topK_pred[1][1][i][0]
if int(pred_0) == int(y_new[i]):
correct_count += 1
fontdict = {'verticalalignment': 'baseline', 'horizontalalignment': 'left', 'color': 'green'}
plt.title("* CORRECT *", fontdict)
else:
fontdict = {'verticalalignment': 'baseline', 'horizontalalignment': 'left', 'color': 'red'}
plt.title("* INCORRECT *", fontdict)
plt.axis('off');
plt.show();
acc_rate = correct_count / len(X_new)
print('Number correct: {}/{} ({:.0%})'.format(correct_count, len(X_new), acc_rate))
Precision & Recall for Original Test Images¶
from sklearn.metrics import confusion_matrix, classification_report
import tensorflow as tf
import numpy as np
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
saver = tf.train.Saver()
saver.restore(sess, MODEL_DIR)
sess = tf.get_default_session()
total_predictions = None
num_examples = len(X_test)
for offset in range(0, num_examples, BATCH_SIZE):
batch_x, batch_y = X_test[offset:offset+BATCH_SIZE], y_test[offset:offset+BATCH_SIZE]
predictions = sess.run(tf.argmax(logits, 1),feed_dict={x: batch_x, y: batch_y, keep_prob: 1.0})
if(total_predictions is None):
total_predictions = predictions
else:
total_predictions = np.hstack([total_predictions,predictions])
mat = confusion_matrix(y_test,total_predictions)
plt.imshow(mat, interpolation='nearest')
report = classification_report(y_test,total_predictions)
print(report)
Worst Performing Images¶
| Class ID | Sign Label | Precision | Recall | F1 Score |
|---|---|---|---|---|
| 27 | Pedestrians | 0.61 | 0.52 | 0.56 |
| 24 | Road narrows on the right | 0.57 | 0.86 | 0.68 |
| 21 | Double curve | 0.73 | 0.73 | 0.73 |
| 37 | Go straight or left | 0.59 | 1.00 | 0.74 |
| 0 | Speed limit (20km/h) | 0.62 | 0.95 | 0.75 |
| 29 | Bicycles crossing | 0.64 | 0.96 | 0.77 |
Precision & Recall for New Images¶
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
saver = tf.train.Saver()
saver.restore(sess, MODEL_DIR)
sess = tf.get_default_session()
total_predictions = None
num_examples = len(X_test)
for offset in range(0, num_examples, BATCH_SIZE):
batch_x, batch_y = X_new_norm[offset:offset+BATCH_SIZE], y_new[offset:offset+BATCH_SIZE]
predictions = sess.run(tf.argmax(logits, 1),feed_dict={x: batch_x, y: batch_y, keep_prob: 1.0})
if(total_predictions is None):
total_predictions = predictions
else:
total_predictions = np.hstack([total_predictions,predictions])
mat = confusion_matrix(y_new,total_predictions)
plt.imshow(mat, interpolation='nearest')
report = classification_report(y_new,total_predictions)
print(report)
Note: Once you have completed all of the code implementations and successfully answered each question above, you may finalize your work by exporting the iPython Notebook as an HTML document. You can do this by using the menu above and navigating to \n", "File -> Download as -> HTML (.html). Include the finished document along with this notebook as your submission.
Step 4 (Optional): Visualize the Neural Network's State with Test Images¶
This Section is not required to complete but acts as an additional excersise for understaning the output of a neural network's weights. While neural networks can be a great learning device they are often referred to as a black box. We can understand what the weights of a neural network look like better by plotting their feature maps. After successfully training your neural network you can see what it's feature maps look like by plotting the output of the network's weight layers in response to a test stimuli image. From these plotted feature maps, it's possible to see what characteristics of an image the network finds interesting. For a sign, maybe the inner network feature maps react with high activation to the sign's boundary outline or to the contrast in the sign's painted symbol.
Provided for you below is the function code that allows you to get the visualization output of any tensorflow weight layer you want. The inputs to the function should be a stimuli image, one used during training or a new one you provided, and then the tensorflow variable name that represents the layer's state during the training process, for instance if you wanted to see what the LeNet lab's feature maps looked like for it's second convolutional layer you could enter conv2 as the tf_activation variable.
For an example of what feature map outputs look like, check out NVIDIA's results in their paper End-to-End Deep Learning for Self-Driving Cars in the section Visualization of internal CNN State. NVIDIA was able to show that their network's inner weights had high activations to road boundary lines by comparing feature maps from an image with a clear path to one without. Try experimenting with a similar test to show that your trained network's weights are looking for interesting features, whether it's looking at differences in feature maps from images with or without a sign, or even what feature maps look like in a trained network vs a completely untrained one on the same sign image.
Your output should look something like this (above)